4 * This file contains the various mmu fetch and update operations.
5 * The most important job they must perform is the mapping between the
6 * domain's pfn and the overall machine mfns.
8 * Xen allows guests to directly update the pagetable, in a controlled
9 * fashion. In other words, the guest modifies the same pagetable
10 * that the CPU actually uses, which eliminates the overhead of having
11 * a separate shadow pagetable.
13 * In order to allow this, it falls on the guest domain to map its
14 * notion of a "physical" pfn - which is just a domain-local linear
15 * address - into a real "machine address" which the CPU's MMU can
18 * A pgd_t/pmd_t/pte_t will typically contain an mfn, and so can be
19 * inserted directly into the pagetable. When creating a new
20 * pte/pmd/pgd, it converts the passed pfn into an mfn. Conversely,
21 * when reading the content back with __(pgd|pmd|pte)_val, it converts
22 * the mfn back into a pfn.
24 * The other constraint is that all pages which make up a pagetable
25 * must be mapped read-only in the guest. This prevents uncontrolled
26 * guest updates to the pagetable. Xen strictly enforces this, and
27 * will disallow any pagetable update which will end up mapping a
28 * pagetable page RW, and will disallow using any writable page as a
31 * Naively, when loading %cr3 with the base of a new pagetable, Xen
32 * would need to validate the whole pagetable before going on.
33 * Naturally, this is quite slow. The solution is to "pin" a
34 * pagetable, which enforces all the constraints on the pagetable even
35 * when it is not actively in use. This menas that Xen can be assured
36 * that it is still valid when you do load it into %cr3, and doesn't
37 * need to revalidate it.
39 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
41 #include <linux/sched.h>
42 #include <linux/highmem.h>
43 #include <linux/debugfs.h>
44 #include <linux/bug.h>
45 #include <linux/vmalloc.h>
46 #include <linux/module.h>
47 #include <linux/gfp.h>
48 #include <linux/memblock.h>
49 #include <linux/seq_file.h>
51 #include <trace/events/xen.h>
53 #include <asm/pgtable.h>
54 #include <asm/tlbflush.h>
55 #include <asm/fixmap.h>
56 #include <asm/mmu_context.h>
57 #include <asm/setup.h>
58 #include <asm/paravirt.h>
60 #include <asm/linkage.h>
66 #include <asm/xen/hypercall.h>
67 #include <asm/xen/hypervisor.h>
71 #include <xen/interface/xen.h>
72 #include <xen/interface/hvm/hvm_op.h>
73 #include <xen/interface/version.h>
74 #include <xen/interface/memory.h>
75 #include <xen/hvc-console.h>
77 #include "multicalls.h"
82 * Protects atomic reservation decrease/increase against concurrent increases.
83 * Also protects non-atomic updates of current_pages and balloon lists.
85 DEFINE_SPINLOCK(xen_reservation_lock);
88 * Identity map, in addition to plain kernel map. This needs to be
89 * large enough to allocate page table pages to allocate the rest.
90 * Each page can map 2MB.
92 #define LEVEL1_IDENT_ENTRIES (PTRS_PER_PTE * 4)
93 static RESERVE_BRK_ARRAY(pte_t, level1_ident_pgt, LEVEL1_IDENT_ENTRIES);
96 /* l3 pud for userspace vsyscall mapping */
97 static pud_t level3_user_vsyscall[PTRS_PER_PUD] __page_aligned_bss;
98 #endif /* CONFIG_X86_64 */
101 * Note about cr3 (pagetable base) values:
103 * xen_cr3 contains the current logical cr3 value; it contains the
104 * last set cr3. This may not be the current effective cr3, because
105 * its update may be being lazily deferred. However, a vcpu looking
106 * at its own cr3 can use this value knowing that it everything will
107 * be self-consistent.
109 * xen_current_cr3 contains the actual vcpu cr3; it is set once the
110 * hypercall to set the vcpu cr3 is complete (so it may be a little
111 * out of date, but it will never be set early). If one vcpu is
112 * looking at another vcpu's cr3 value, it should use this variable.
114 DEFINE_PER_CPU(unsigned long, xen_cr3); /* cr3 stored as physaddr */
115 DEFINE_PER_CPU(unsigned long, xen_current_cr3); /* actual vcpu cr3 */
119 * Just beyond the highest usermode address. STACK_TOP_MAX has a
120 * redzone above it, so round it up to a PGD boundary.
122 #define USER_LIMIT ((STACK_TOP_MAX + PGDIR_SIZE - 1) & PGDIR_MASK)
124 unsigned long arbitrary_virt_to_mfn(void *vaddr)
126 xmaddr_t maddr = arbitrary_virt_to_machine(vaddr);
128 return PFN_DOWN(maddr.maddr);
131 xmaddr_t arbitrary_virt_to_machine(void *vaddr)
133 unsigned long address = (unsigned long)vaddr;
139 * if the PFN is in the linear mapped vaddr range, we can just use
140 * the (quick) virt_to_machine() p2m lookup
142 if (virt_addr_valid(vaddr))
143 return virt_to_machine(vaddr);
145 /* otherwise we have to do a (slower) full page-table walk */
147 pte = lookup_address(address, &level);
149 offset = address & ~PAGE_MASK;
150 return XMADDR(((phys_addr_t)pte_mfn(*pte) << PAGE_SHIFT) + offset);
152 EXPORT_SYMBOL_GPL(arbitrary_virt_to_machine);
154 void make_lowmem_page_readonly(void *vaddr)
157 unsigned long address = (unsigned long)vaddr;
160 pte = lookup_address(address, &level);
162 return; /* vaddr missing */
164 ptev = pte_wrprotect(*pte);
166 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
170 void make_lowmem_page_readwrite(void *vaddr)
173 unsigned long address = (unsigned long)vaddr;
176 pte = lookup_address(address, &level);
178 return; /* vaddr missing */
180 ptev = pte_mkwrite(*pte);
182 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
187 static bool xen_page_pinned(void *ptr)
189 struct page *page = virt_to_page(ptr);
191 return PagePinned(page);
194 void xen_set_domain_pte(pte_t *ptep, pte_t pteval, unsigned domid)
196 struct multicall_space mcs;
197 struct mmu_update *u;
199 trace_xen_mmu_set_domain_pte(ptep, pteval, domid);
201 mcs = xen_mc_entry(sizeof(*u));
204 /* ptep might be kmapped when using 32-bit HIGHPTE */
205 u->ptr = virt_to_machine(ptep).maddr;
206 u->val = pte_val_ma(pteval);
208 MULTI_mmu_update(mcs.mc, mcs.args, 1, NULL, domid);
210 xen_mc_issue(PARAVIRT_LAZY_MMU);
212 EXPORT_SYMBOL_GPL(xen_set_domain_pte);
214 static void xen_extend_mmu_update(const struct mmu_update *update)
216 struct multicall_space mcs;
217 struct mmu_update *u;
219 mcs = xen_mc_extend_args(__HYPERVISOR_mmu_update, sizeof(*u));
221 if (mcs.mc != NULL) {
224 mcs = __xen_mc_entry(sizeof(*u));
225 MULTI_mmu_update(mcs.mc, mcs.args, 1, NULL, DOMID_SELF);
232 static void xen_extend_mmuext_op(const struct mmuext_op *op)
234 struct multicall_space mcs;
237 mcs = xen_mc_extend_args(__HYPERVISOR_mmuext_op, sizeof(*u));
239 if (mcs.mc != NULL) {
242 mcs = __xen_mc_entry(sizeof(*u));
243 MULTI_mmuext_op(mcs.mc, mcs.args, 1, NULL, DOMID_SELF);
250 static void xen_set_pmd_hyper(pmd_t *ptr, pmd_t val)
258 /* ptr may be ioremapped for 64-bit pagetable setup */
259 u.ptr = arbitrary_virt_to_machine(ptr).maddr;
260 u.val = pmd_val_ma(val);
261 xen_extend_mmu_update(&u);
263 xen_mc_issue(PARAVIRT_LAZY_MMU);
268 static void xen_set_pmd(pmd_t *ptr, pmd_t val)
270 trace_xen_mmu_set_pmd(ptr, val);
272 /* If page is not pinned, we can just update the entry
274 if (!xen_page_pinned(ptr)) {
279 xen_set_pmd_hyper(ptr, val);
283 * Associate a virtual page frame with a given physical page frame
284 * and protection flags for that frame.
286 void set_pte_mfn(unsigned long vaddr, unsigned long mfn, pgprot_t flags)
288 set_pte_vaddr(vaddr, mfn_pte(mfn, flags));
291 static bool xen_batched_set_pte(pte_t *ptep, pte_t pteval)
295 if (paravirt_get_lazy_mode() != PARAVIRT_LAZY_MMU)
300 u.ptr = virt_to_machine(ptep).maddr | MMU_NORMAL_PT_UPDATE;
301 u.val = pte_val_ma(pteval);
302 xen_extend_mmu_update(&u);
304 xen_mc_issue(PARAVIRT_LAZY_MMU);
309 static inline void __xen_set_pte(pte_t *ptep, pte_t pteval)
311 if (!xen_batched_set_pte(ptep, pteval)) {
313 * Could call native_set_pte() here and trap and
314 * emulate the PTE write but with 32-bit guests this
315 * needs two traps (one for each of the two 32-bit
316 * words in the PTE) so do one hypercall directly
321 u.ptr = virt_to_machine(ptep).maddr | MMU_NORMAL_PT_UPDATE;
322 u.val = pte_val_ma(pteval);
323 HYPERVISOR_mmu_update(&u, 1, NULL, DOMID_SELF);
327 static void xen_set_pte(pte_t *ptep, pte_t pteval)
329 trace_xen_mmu_set_pte(ptep, pteval);
330 __xen_set_pte(ptep, pteval);
333 static void xen_set_pte_at(struct mm_struct *mm, unsigned long addr,
334 pte_t *ptep, pte_t pteval)
336 trace_xen_mmu_set_pte_at(mm, addr, ptep, pteval);
337 __xen_set_pte(ptep, pteval);
340 pte_t xen_ptep_modify_prot_start(struct mm_struct *mm,
341 unsigned long addr, pte_t *ptep)
343 /* Just return the pte as-is. We preserve the bits on commit */
344 trace_xen_mmu_ptep_modify_prot_start(mm, addr, ptep, *ptep);
348 void xen_ptep_modify_prot_commit(struct mm_struct *mm, unsigned long addr,
349 pte_t *ptep, pte_t pte)
353 trace_xen_mmu_ptep_modify_prot_commit(mm, addr, ptep, pte);
356 u.ptr = virt_to_machine(ptep).maddr | MMU_PT_UPDATE_PRESERVE_AD;
357 u.val = pte_val_ma(pte);
358 xen_extend_mmu_update(&u);
360 xen_mc_issue(PARAVIRT_LAZY_MMU);
363 /* Assume pteval_t is equivalent to all the other *val_t types. */
364 static pteval_t pte_mfn_to_pfn(pteval_t val)
366 if (val & _PAGE_PRESENT) {
367 unsigned long mfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
368 unsigned long pfn = mfn_to_pfn(mfn);
370 pteval_t flags = val & PTE_FLAGS_MASK;
371 if (unlikely(pfn == ~0))
372 val = flags & ~_PAGE_PRESENT;
374 val = ((pteval_t)pfn << PAGE_SHIFT) | flags;
380 static pteval_t pte_pfn_to_mfn(pteval_t val)
382 if (val & _PAGE_PRESENT) {
383 unsigned long pfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
384 pteval_t flags = val & PTE_FLAGS_MASK;
387 if (!xen_feature(XENFEAT_auto_translated_physmap))
388 mfn = get_phys_to_machine(pfn);
392 * If there's no mfn for the pfn, then just create an
393 * empty non-present pte. Unfortunately this loses
394 * information about the original pfn, so
395 * pte_mfn_to_pfn is asymmetric.
397 if (unlikely(mfn == INVALID_P2M_ENTRY)) {
402 * Paramount to do this test _after_ the
403 * INVALID_P2M_ENTRY as INVALID_P2M_ENTRY &
404 * IDENTITY_FRAME_BIT resolves to true.
406 mfn &= ~FOREIGN_FRAME_BIT;
407 if (mfn & IDENTITY_FRAME_BIT) {
408 mfn &= ~IDENTITY_FRAME_BIT;
409 flags |= _PAGE_IOMAP;
412 val = ((pteval_t)mfn << PAGE_SHIFT) | flags;
418 static pteval_t iomap_pte(pteval_t val)
420 if (val & _PAGE_PRESENT) {
421 unsigned long pfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
422 pteval_t flags = val & PTE_FLAGS_MASK;
424 /* We assume the pte frame number is a MFN, so
425 just use it as-is. */
426 val = ((pteval_t)pfn << PAGE_SHIFT) | flags;
432 static pteval_t xen_pte_val(pte_t pte)
434 pteval_t pteval = pte.pte;
436 /* If this is a WC pte, convert back from Xen WC to Linux WC */
437 if ((pteval & (_PAGE_PAT | _PAGE_PCD | _PAGE_PWT)) == _PAGE_PAT) {
438 WARN_ON(!pat_enabled);
439 pteval = (pteval & ~_PAGE_PAT) | _PAGE_PWT;
442 if (xen_initial_domain() && (pteval & _PAGE_IOMAP))
445 return pte_mfn_to_pfn(pteval);
447 PV_CALLEE_SAVE_REGS_THUNK(xen_pte_val);
449 static pgdval_t xen_pgd_val(pgd_t pgd)
451 return pte_mfn_to_pfn(pgd.pgd);
453 PV_CALLEE_SAVE_REGS_THUNK(xen_pgd_val);
456 * Xen's PAT setup is part of its ABI, though I assume entries 6 & 7
457 * are reserved for now, to correspond to the Intel-reserved PAT
460 * We expect Linux's PAT set as follows:
462 * Idx PTE flags Linux Xen Default
469 * 6 PAT PCD UC- UC UC-
470 * 7 PAT PCD PWT UC UC UC
473 void xen_set_pat(u64 pat)
475 /* We expect Linux to use a PAT setting of
476 * UC UC- WC WB (ignoring the PAT flag) */
477 WARN_ON(pat != 0x0007010600070106ull);
480 static pte_t xen_make_pte(pteval_t pte)
482 phys_addr_t addr = (pte & PTE_PFN_MASK);
484 /* If Linux is trying to set a WC pte, then map to the Xen WC.
485 * If _PAGE_PAT is set, then it probably means it is really
486 * _PAGE_PSE, so avoid fiddling with the PAT mapping and hope
487 * things work out OK...
489 * (We should never see kernel mappings with _PAGE_PSE set,
490 * but we could see hugetlbfs mappings, I think.).
492 if (pat_enabled && !WARN_ON(pte & _PAGE_PAT)) {
493 if ((pte & (_PAGE_PCD | _PAGE_PWT)) == _PAGE_PWT)
494 pte = (pte & ~(_PAGE_PCD | _PAGE_PWT)) | _PAGE_PAT;
498 * Unprivileged domains are allowed to do IOMAPpings for
499 * PCI passthrough, but not map ISA space. The ISA
500 * mappings are just dummy local mappings to keep other
501 * parts of the kernel happy.
503 if (unlikely(pte & _PAGE_IOMAP) &&
504 (xen_initial_domain() || addr >= ISA_END_ADDRESS)) {
505 pte = iomap_pte(pte);
508 pte = pte_pfn_to_mfn(pte);
511 return native_make_pte(pte);
513 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pte);
515 static pgd_t xen_make_pgd(pgdval_t pgd)
517 pgd = pte_pfn_to_mfn(pgd);
518 return native_make_pgd(pgd);
520 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pgd);
522 static pmdval_t xen_pmd_val(pmd_t pmd)
524 return pte_mfn_to_pfn(pmd.pmd);
526 PV_CALLEE_SAVE_REGS_THUNK(xen_pmd_val);
528 static void xen_set_pud_hyper(pud_t *ptr, pud_t val)
536 /* ptr may be ioremapped for 64-bit pagetable setup */
537 u.ptr = arbitrary_virt_to_machine(ptr).maddr;
538 u.val = pud_val_ma(val);
539 xen_extend_mmu_update(&u);
541 xen_mc_issue(PARAVIRT_LAZY_MMU);
546 static void xen_set_pud(pud_t *ptr, pud_t val)
548 trace_xen_mmu_set_pud(ptr, val);
550 /* If page is not pinned, we can just update the entry
552 if (!xen_page_pinned(ptr)) {
557 xen_set_pud_hyper(ptr, val);
560 #ifdef CONFIG_X86_PAE
561 static void xen_set_pte_atomic(pte_t *ptep, pte_t pte)
563 trace_xen_mmu_set_pte_atomic(ptep, pte);
564 set_64bit((u64 *)ptep, native_pte_val(pte));
567 static void xen_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
569 trace_xen_mmu_pte_clear(mm, addr, ptep);
570 if (!xen_batched_set_pte(ptep, native_make_pte(0)))
571 native_pte_clear(mm, addr, ptep);
574 static void xen_pmd_clear(pmd_t *pmdp)
576 trace_xen_mmu_pmd_clear(pmdp);
577 set_pmd(pmdp, __pmd(0));
579 #endif /* CONFIG_X86_PAE */
581 static pmd_t xen_make_pmd(pmdval_t pmd)
583 pmd = pte_pfn_to_mfn(pmd);
584 return native_make_pmd(pmd);
586 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pmd);
588 #if PAGETABLE_LEVELS == 4
589 static pudval_t xen_pud_val(pud_t pud)
591 return pte_mfn_to_pfn(pud.pud);
593 PV_CALLEE_SAVE_REGS_THUNK(xen_pud_val);
595 static pud_t xen_make_pud(pudval_t pud)
597 pud = pte_pfn_to_mfn(pud);
599 return native_make_pud(pud);
601 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pud);
603 static pgd_t *xen_get_user_pgd(pgd_t *pgd)
605 pgd_t *pgd_page = (pgd_t *)(((unsigned long)pgd) & PAGE_MASK);
606 unsigned offset = pgd - pgd_page;
607 pgd_t *user_ptr = NULL;
609 if (offset < pgd_index(USER_LIMIT)) {
610 struct page *page = virt_to_page(pgd_page);
611 user_ptr = (pgd_t *)page->private;
619 static void __xen_set_pgd_hyper(pgd_t *ptr, pgd_t val)
623 u.ptr = virt_to_machine(ptr).maddr;
624 u.val = pgd_val_ma(val);
625 xen_extend_mmu_update(&u);
629 * Raw hypercall-based set_pgd, intended for in early boot before
630 * there's a page structure. This implies:
631 * 1. The only existing pagetable is the kernel's
632 * 2. It is always pinned
633 * 3. It has no user pagetable attached to it
635 static void __init xen_set_pgd_hyper(pgd_t *ptr, pgd_t val)
641 __xen_set_pgd_hyper(ptr, val);
643 xen_mc_issue(PARAVIRT_LAZY_MMU);
648 static void xen_set_pgd(pgd_t *ptr, pgd_t val)
650 pgd_t *user_ptr = xen_get_user_pgd(ptr);
652 trace_xen_mmu_set_pgd(ptr, user_ptr, val);
654 /* If page is not pinned, we can just update the entry
656 if (!xen_page_pinned(ptr)) {
659 WARN_ON(xen_page_pinned(user_ptr));
665 /* If it's pinned, then we can at least batch the kernel and
666 user updates together. */
669 __xen_set_pgd_hyper(ptr, val);
671 __xen_set_pgd_hyper(user_ptr, val);
673 xen_mc_issue(PARAVIRT_LAZY_MMU);
675 #endif /* PAGETABLE_LEVELS == 4 */
678 * (Yet another) pagetable walker. This one is intended for pinning a
679 * pagetable. This means that it walks a pagetable and calls the
680 * callback function on each page it finds making up the page table,
681 * at every level. It walks the entire pagetable, but it only bothers
682 * pinning pte pages which are below limit. In the normal case this
683 * will be STACK_TOP_MAX, but at boot we need to pin up to
686 * For 32-bit the important bit is that we don't pin beyond there,
687 * because then we start getting into Xen's ptes.
689 * For 64-bit, we must skip the Xen hole in the middle of the address
690 * space, just after the big x86-64 virtual hole.
692 static int __xen_pgd_walk(struct mm_struct *mm, pgd_t *pgd,
693 int (*func)(struct mm_struct *mm, struct page *,
698 unsigned hole_low, hole_high;
699 unsigned pgdidx_limit, pudidx_limit, pmdidx_limit;
700 unsigned pgdidx, pudidx, pmdidx;
702 /* The limit is the last byte to be touched */
704 BUG_ON(limit >= FIXADDR_TOP);
706 if (xen_feature(XENFEAT_auto_translated_physmap))
710 * 64-bit has a great big hole in the middle of the address
711 * space, which contains the Xen mappings. On 32-bit these
712 * will end up making a zero-sized hole and so is a no-op.
714 hole_low = pgd_index(USER_LIMIT);
715 hole_high = pgd_index(PAGE_OFFSET);
717 pgdidx_limit = pgd_index(limit);
719 pudidx_limit = pud_index(limit);
724 pmdidx_limit = pmd_index(limit);
729 for (pgdidx = 0; pgdidx <= pgdidx_limit; pgdidx++) {
732 if (pgdidx >= hole_low && pgdidx < hole_high)
735 if (!pgd_val(pgd[pgdidx]))
738 pud = pud_offset(&pgd[pgdidx], 0);
740 if (PTRS_PER_PUD > 1) /* not folded */
741 flush |= (*func)(mm, virt_to_page(pud), PT_PUD);
743 for (pudidx = 0; pudidx < PTRS_PER_PUD; pudidx++) {
746 if (pgdidx == pgdidx_limit &&
747 pudidx > pudidx_limit)
750 if (pud_none(pud[pudidx]))
753 pmd = pmd_offset(&pud[pudidx], 0);
755 if (PTRS_PER_PMD > 1) /* not folded */
756 flush |= (*func)(mm, virt_to_page(pmd), PT_PMD);
758 for (pmdidx = 0; pmdidx < PTRS_PER_PMD; pmdidx++) {
761 if (pgdidx == pgdidx_limit &&
762 pudidx == pudidx_limit &&
763 pmdidx > pmdidx_limit)
766 if (pmd_none(pmd[pmdidx]))
769 pte = pmd_page(pmd[pmdidx]);
770 flush |= (*func)(mm, pte, PT_PTE);
776 /* Do the top level last, so that the callbacks can use it as
777 a cue to do final things like tlb flushes. */
778 flush |= (*func)(mm, virt_to_page(pgd), PT_PGD);
783 static int xen_pgd_walk(struct mm_struct *mm,
784 int (*func)(struct mm_struct *mm, struct page *,
788 return __xen_pgd_walk(mm, mm->pgd, func, limit);
791 /* If we're using split pte locks, then take the page's lock and
792 return a pointer to it. Otherwise return NULL. */
793 static spinlock_t *xen_pte_lock(struct page *page, struct mm_struct *mm)
795 spinlock_t *ptl = NULL;
797 #if USE_SPLIT_PTLOCKS
798 ptl = __pte_lockptr(page);
799 spin_lock_nest_lock(ptl, &mm->page_table_lock);
805 static void xen_pte_unlock(void *v)
811 static void xen_do_pin(unsigned level, unsigned long pfn)
816 op.arg1.mfn = pfn_to_mfn(pfn);
818 xen_extend_mmuext_op(&op);
821 static int xen_pin_page(struct mm_struct *mm, struct page *page,
824 unsigned pgfl = TestSetPagePinned(page);
828 flush = 0; /* already pinned */
829 else if (PageHighMem(page))
830 /* kmaps need flushing if we found an unpinned
834 void *pt = lowmem_page_address(page);
835 unsigned long pfn = page_to_pfn(page);
836 struct multicall_space mcs = __xen_mc_entry(0);
842 * We need to hold the pagetable lock between the time
843 * we make the pagetable RO and when we actually pin
844 * it. If we don't, then other users may come in and
845 * attempt to update the pagetable by writing it,
846 * which will fail because the memory is RO but not
847 * pinned, so Xen won't do the trap'n'emulate.
849 * If we're using split pte locks, we can't hold the
850 * entire pagetable's worth of locks during the
851 * traverse, because we may wrap the preempt count (8
852 * bits). The solution is to mark RO and pin each PTE
853 * page while holding the lock. This means the number
854 * of locks we end up holding is never more than a
855 * batch size (~32 entries, at present).
857 * If we're not using split pte locks, we needn't pin
858 * the PTE pages independently, because we're
859 * protected by the overall pagetable lock.
863 ptl = xen_pte_lock(page, mm);
865 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
866 pfn_pte(pfn, PAGE_KERNEL_RO),
867 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
870 xen_do_pin(MMUEXT_PIN_L1_TABLE, pfn);
872 /* Queue a deferred unlock for when this batch
874 xen_mc_callback(xen_pte_unlock, ptl);
881 /* This is called just after a mm has been created, but it has not
882 been used yet. We need to make sure that its pagetable is all
883 read-only, and can be pinned. */
884 static void __xen_pgd_pin(struct mm_struct *mm, pgd_t *pgd)
886 trace_xen_mmu_pgd_pin(mm, pgd);
890 if (__xen_pgd_walk(mm, pgd, xen_pin_page, USER_LIMIT)) {
891 /* re-enable interrupts for flushing */
901 pgd_t *user_pgd = xen_get_user_pgd(pgd);
903 xen_do_pin(MMUEXT_PIN_L4_TABLE, PFN_DOWN(__pa(pgd)));
906 xen_pin_page(mm, virt_to_page(user_pgd), PT_PGD);
907 xen_do_pin(MMUEXT_PIN_L4_TABLE,
908 PFN_DOWN(__pa(user_pgd)));
911 #else /* CONFIG_X86_32 */
912 #ifdef CONFIG_X86_PAE
913 /* Need to make sure unshared kernel PMD is pinnable */
914 xen_pin_page(mm, pgd_page(pgd[pgd_index(TASK_SIZE)]),
917 xen_do_pin(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(pgd)));
918 #endif /* CONFIG_X86_64 */
922 static void xen_pgd_pin(struct mm_struct *mm)
924 __xen_pgd_pin(mm, mm->pgd);
928 * On save, we need to pin all pagetables to make sure they get their
929 * mfns turned into pfns. Search the list for any unpinned pgds and pin
930 * them (unpinned pgds are not currently in use, probably because the
931 * process is under construction or destruction).
933 * Expected to be called in stop_machine() ("equivalent to taking
934 * every spinlock in the system"), so the locking doesn't really
935 * matter all that much.
937 void xen_mm_pin_all(void)
941 spin_lock(&pgd_lock);
943 list_for_each_entry(page, &pgd_list, lru) {
944 if (!PagePinned(page)) {
945 __xen_pgd_pin(&init_mm, (pgd_t *)page_address(page));
946 SetPageSavePinned(page);
950 spin_unlock(&pgd_lock);
954 * The init_mm pagetable is really pinned as soon as its created, but
955 * that's before we have page structures to store the bits. So do all
956 * the book-keeping now.
958 static int __init xen_mark_pinned(struct mm_struct *mm, struct page *page,
965 static void __init xen_mark_init_mm_pinned(void)
967 xen_pgd_walk(&init_mm, xen_mark_pinned, FIXADDR_TOP);
970 static int xen_unpin_page(struct mm_struct *mm, struct page *page,
973 unsigned pgfl = TestClearPagePinned(page);
975 if (pgfl && !PageHighMem(page)) {
976 void *pt = lowmem_page_address(page);
977 unsigned long pfn = page_to_pfn(page);
978 spinlock_t *ptl = NULL;
979 struct multicall_space mcs;
982 * Do the converse to pin_page. If we're using split
983 * pte locks, we must be holding the lock for while
984 * the pte page is unpinned but still RO to prevent
985 * concurrent updates from seeing it in this
986 * partially-pinned state.
988 if (level == PT_PTE) {
989 ptl = xen_pte_lock(page, mm);
992 xen_do_pin(MMUEXT_UNPIN_TABLE, pfn);
995 mcs = __xen_mc_entry(0);
997 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
998 pfn_pte(pfn, PAGE_KERNEL),
999 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
1002 /* unlock when batch completed */
1003 xen_mc_callback(xen_pte_unlock, ptl);
1007 return 0; /* never need to flush on unpin */
1010 /* Release a pagetables pages back as normal RW */
1011 static void __xen_pgd_unpin(struct mm_struct *mm, pgd_t *pgd)
1013 trace_xen_mmu_pgd_unpin(mm, pgd);
1017 xen_do_pin(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1019 #ifdef CONFIG_X86_64
1021 pgd_t *user_pgd = xen_get_user_pgd(pgd);
1024 xen_do_pin(MMUEXT_UNPIN_TABLE,
1025 PFN_DOWN(__pa(user_pgd)));
1026 xen_unpin_page(mm, virt_to_page(user_pgd), PT_PGD);
1031 #ifdef CONFIG_X86_PAE
1032 /* Need to make sure unshared kernel PMD is unpinned */
1033 xen_unpin_page(mm, pgd_page(pgd[pgd_index(TASK_SIZE)]),
1037 __xen_pgd_walk(mm, pgd, xen_unpin_page, USER_LIMIT);
1042 static void xen_pgd_unpin(struct mm_struct *mm)
1044 __xen_pgd_unpin(mm, mm->pgd);
1048 * On resume, undo any pinning done at save, so that the rest of the
1049 * kernel doesn't see any unexpected pinned pagetables.
1051 void xen_mm_unpin_all(void)
1055 spin_lock(&pgd_lock);
1057 list_for_each_entry(page, &pgd_list, lru) {
1058 if (PageSavePinned(page)) {
1059 BUG_ON(!PagePinned(page));
1060 __xen_pgd_unpin(&init_mm, (pgd_t *)page_address(page));
1061 ClearPageSavePinned(page);
1065 spin_unlock(&pgd_lock);
1068 static void xen_activate_mm(struct mm_struct *prev, struct mm_struct *next)
1070 spin_lock(&next->page_table_lock);
1072 spin_unlock(&next->page_table_lock);
1075 static void xen_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
1077 spin_lock(&mm->page_table_lock);
1079 spin_unlock(&mm->page_table_lock);
1084 /* Another cpu may still have their %cr3 pointing at the pagetable, so
1085 we need to repoint it somewhere else before we can unpin it. */
1086 static void drop_other_mm_ref(void *info)
1088 struct mm_struct *mm = info;
1089 struct mm_struct *active_mm;
1091 active_mm = this_cpu_read(cpu_tlbstate.active_mm);
1093 if (active_mm == mm && this_cpu_read(cpu_tlbstate.state) != TLBSTATE_OK)
1094 leave_mm(smp_processor_id());
1096 /* If this cpu still has a stale cr3 reference, then make sure
1097 it has been flushed. */
1098 if (this_cpu_read(xen_current_cr3) == __pa(mm->pgd))
1099 load_cr3(swapper_pg_dir);
1102 static void xen_drop_mm_ref(struct mm_struct *mm)
1107 if (current->active_mm == mm) {
1108 if (current->mm == mm)
1109 load_cr3(swapper_pg_dir);
1111 leave_mm(smp_processor_id());
1114 /* Get the "official" set of cpus referring to our pagetable. */
1115 if (!alloc_cpumask_var(&mask, GFP_ATOMIC)) {
1116 for_each_online_cpu(cpu) {
1117 if (!cpumask_test_cpu(cpu, mm_cpumask(mm))
1118 && per_cpu(xen_current_cr3, cpu) != __pa(mm->pgd))
1120 smp_call_function_single(cpu, drop_other_mm_ref, mm, 1);
1124 cpumask_copy(mask, mm_cpumask(mm));
1126 /* It's possible that a vcpu may have a stale reference to our
1127 cr3, because its in lazy mode, and it hasn't yet flushed
1128 its set of pending hypercalls yet. In this case, we can
1129 look at its actual current cr3 value, and force it to flush
1131 for_each_online_cpu(cpu) {
1132 if (per_cpu(xen_current_cr3, cpu) == __pa(mm->pgd))
1133 cpumask_set_cpu(cpu, mask);
1136 if (!cpumask_empty(mask))
1137 smp_call_function_many(mask, drop_other_mm_ref, mm, 1);
1138 free_cpumask_var(mask);
1141 static void xen_drop_mm_ref(struct mm_struct *mm)
1143 if (current->active_mm == mm)
1144 load_cr3(swapper_pg_dir);
1149 * While a process runs, Xen pins its pagetables, which means that the
1150 * hypervisor forces it to be read-only, and it controls all updates
1151 * to it. This means that all pagetable updates have to go via the
1152 * hypervisor, which is moderately expensive.
1154 * Since we're pulling the pagetable down, we switch to use init_mm,
1155 * unpin old process pagetable and mark it all read-write, which
1156 * allows further operations on it to be simple memory accesses.
1158 * The only subtle point is that another CPU may be still using the
1159 * pagetable because of lazy tlb flushing. This means we need need to
1160 * switch all CPUs off this pagetable before we can unpin it.
1162 static void xen_exit_mmap(struct mm_struct *mm)
1164 get_cpu(); /* make sure we don't move around */
1165 xen_drop_mm_ref(mm);
1168 spin_lock(&mm->page_table_lock);
1170 /* pgd may not be pinned in the error exit path of execve */
1171 if (xen_page_pinned(mm->pgd))
1174 spin_unlock(&mm->page_table_lock);
1177 static void __init xen_pagetable_setup_start(pgd_t *base)
1181 static __init void xen_mapping_pagetable_reserve(u64 start, u64 end)
1183 /* reserve the range used */
1184 native_pagetable_reserve(start, end);
1186 /* set as RW the rest */
1187 printk(KERN_DEBUG "xen: setting RW the range %llx - %llx\n", end,
1188 PFN_PHYS(pgt_buf_top));
1189 while (end < PFN_PHYS(pgt_buf_top)) {
1190 make_lowmem_page_readwrite(__va(end));
1195 static void xen_post_allocator_init(void);
1197 static void __init xen_pagetable_setup_done(pgd_t *base)
1199 xen_setup_shared_info();
1200 xen_post_allocator_init();
1203 static void xen_write_cr2(unsigned long cr2)
1205 this_cpu_read(xen_vcpu)->arch.cr2 = cr2;
1208 static unsigned long xen_read_cr2(void)
1210 return this_cpu_read(xen_vcpu)->arch.cr2;
1213 unsigned long xen_read_cr2_direct(void)
1215 return this_cpu_read(xen_vcpu_info.arch.cr2);
1218 static void xen_flush_tlb(void)
1220 struct mmuext_op *op;
1221 struct multicall_space mcs;
1223 trace_xen_mmu_flush_tlb(0);
1227 mcs = xen_mc_entry(sizeof(*op));
1230 op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
1231 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1233 xen_mc_issue(PARAVIRT_LAZY_MMU);
1238 static void xen_flush_tlb_single(unsigned long addr)
1240 struct mmuext_op *op;
1241 struct multicall_space mcs;
1243 trace_xen_mmu_flush_tlb_single(addr);
1247 mcs = xen_mc_entry(sizeof(*op));
1249 op->cmd = MMUEXT_INVLPG_LOCAL;
1250 op->arg1.linear_addr = addr & PAGE_MASK;
1251 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1253 xen_mc_issue(PARAVIRT_LAZY_MMU);
1258 static void xen_flush_tlb_others(const struct cpumask *cpus,
1259 struct mm_struct *mm, unsigned long va)
1262 struct mmuext_op op;
1264 DECLARE_BITMAP(mask, num_processors);
1266 DECLARE_BITMAP(mask, NR_CPUS);
1269 struct multicall_space mcs;
1271 trace_xen_mmu_flush_tlb_others(cpus, mm, va);
1273 if (cpumask_empty(cpus))
1274 return; /* nothing to do */
1276 mcs = xen_mc_entry(sizeof(*args));
1278 args->op.arg2.vcpumask = to_cpumask(args->mask);
1280 /* Remove us, and any offline CPUS. */
1281 cpumask_and(to_cpumask(args->mask), cpus, cpu_online_mask);
1282 cpumask_clear_cpu(smp_processor_id(), to_cpumask(args->mask));
1284 if (va == TLB_FLUSH_ALL) {
1285 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
1287 args->op.cmd = MMUEXT_INVLPG_MULTI;
1288 args->op.arg1.linear_addr = va;
1291 MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
1293 xen_mc_issue(PARAVIRT_LAZY_MMU);
1296 static unsigned long xen_read_cr3(void)
1298 return this_cpu_read(xen_cr3);
1301 static void set_current_cr3(void *v)
1303 this_cpu_write(xen_current_cr3, (unsigned long)v);
1306 static void __xen_write_cr3(bool kernel, unsigned long cr3)
1308 struct mmuext_op op;
1311 trace_xen_mmu_write_cr3(kernel, cr3);
1314 mfn = pfn_to_mfn(PFN_DOWN(cr3));
1318 WARN_ON(mfn == 0 && kernel);
1320 op.cmd = kernel ? MMUEXT_NEW_BASEPTR : MMUEXT_NEW_USER_BASEPTR;
1323 xen_extend_mmuext_op(&op);
1326 this_cpu_write(xen_cr3, cr3);
1328 /* Update xen_current_cr3 once the batch has actually
1330 xen_mc_callback(set_current_cr3, (void *)cr3);
1334 static void xen_write_cr3(unsigned long cr3)
1336 BUG_ON(preemptible());
1338 xen_mc_batch(); /* disables interrupts */
1340 /* Update while interrupts are disabled, so its atomic with
1342 this_cpu_write(xen_cr3, cr3);
1344 __xen_write_cr3(true, cr3);
1346 #ifdef CONFIG_X86_64
1348 pgd_t *user_pgd = xen_get_user_pgd(__va(cr3));
1350 __xen_write_cr3(false, __pa(user_pgd));
1352 __xen_write_cr3(false, 0);
1356 xen_mc_issue(PARAVIRT_LAZY_CPU); /* interrupts restored */
1359 static int xen_pgd_alloc(struct mm_struct *mm)
1361 pgd_t *pgd = mm->pgd;
1364 BUG_ON(PagePinned(virt_to_page(pgd)));
1366 #ifdef CONFIG_X86_64
1368 struct page *page = virt_to_page(pgd);
1371 BUG_ON(page->private != 0);
1375 user_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1376 page->private = (unsigned long)user_pgd;
1378 if (user_pgd != NULL) {
1379 user_pgd[pgd_index(VSYSCALL_START)] =
1380 __pgd(__pa(level3_user_vsyscall) | _PAGE_TABLE);
1384 BUG_ON(PagePinned(virt_to_page(xen_get_user_pgd(pgd))));
1391 static void xen_pgd_free(struct mm_struct *mm, pgd_t *pgd)
1393 #ifdef CONFIG_X86_64
1394 pgd_t *user_pgd = xen_get_user_pgd(pgd);
1397 free_page((unsigned long)user_pgd);
1401 #ifdef CONFIG_X86_32
1402 static pte_t __init mask_rw_pte(pte_t *ptep, pte_t pte)
1404 /* If there's an existing pte, then don't allow _PAGE_RW to be set */
1405 if (pte_val_ma(*ptep) & _PAGE_PRESENT)
1406 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
1411 #else /* CONFIG_X86_64 */
1412 static pte_t __init mask_rw_pte(pte_t *ptep, pte_t pte)
1414 unsigned long pfn = pte_pfn(pte);
1417 * If the new pfn is within the range of the newly allocated
1418 * kernel pagetable, and it isn't being mapped into an
1419 * early_ioremap fixmap slot as a freshly allocated page, make sure
1422 if (((!is_early_ioremap_ptep(ptep) &&
1423 pfn >= pgt_buf_start && pfn < pgt_buf_top)) ||
1424 (is_early_ioremap_ptep(ptep) && pfn != (pgt_buf_end - 1)))
1425 pte = pte_wrprotect(pte);
1429 #endif /* CONFIG_X86_64 */
1432 * Init-time set_pte while constructing initial pagetables, which
1433 * doesn't allow RO page table pages to be remapped RW.
1435 * Many of these PTE updates are done on unpinned and writable pages
1436 * and doing a hypercall for these is unnecessary and expensive. At
1437 * this point it is not possible to tell if a page is pinned or not,
1438 * so always write the PTE directly and rely on Xen trapping and
1439 * emulating any updates as necessary.
1441 static void __init xen_set_pte_init(pte_t *ptep, pte_t pte)
1443 pte = mask_rw_pte(ptep, pte);
1445 native_set_pte(ptep, pte);
1448 static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
1450 struct mmuext_op op;
1452 op.arg1.mfn = pfn_to_mfn(pfn);
1453 if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
1457 /* Early in boot, while setting up the initial pagetable, assume
1458 everything is pinned. */
1459 static void __init xen_alloc_pte_init(struct mm_struct *mm, unsigned long pfn)
1461 #ifdef CONFIG_FLATMEM
1462 BUG_ON(mem_map); /* should only be used early */
1464 make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
1465 pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
1468 /* Used for pmd and pud */
1469 static void __init xen_alloc_pmd_init(struct mm_struct *mm, unsigned long pfn)
1471 #ifdef CONFIG_FLATMEM
1472 BUG_ON(mem_map); /* should only be used early */
1474 make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
1477 /* Early release_pte assumes that all pts are pinned, since there's
1478 only init_mm and anything attached to that is pinned. */
1479 static void __init xen_release_pte_init(unsigned long pfn)
1481 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
1482 make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1485 static void __init xen_release_pmd_init(unsigned long pfn)
1487 make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1490 static inline void __pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
1492 struct multicall_space mcs;
1493 struct mmuext_op *op;
1495 mcs = __xen_mc_entry(sizeof(*op));
1498 op->arg1.mfn = pfn_to_mfn(pfn);
1500 MULTI_mmuext_op(mcs.mc, mcs.args, 1, NULL, DOMID_SELF);
1503 static inline void __set_pfn_prot(unsigned long pfn, pgprot_t prot)
1505 struct multicall_space mcs;
1506 unsigned long addr = (unsigned long)__va(pfn << PAGE_SHIFT);
1508 mcs = __xen_mc_entry(0);
1509 MULTI_update_va_mapping(mcs.mc, (unsigned long)addr,
1510 pfn_pte(pfn, prot), 0);
1513 /* This needs to make sure the new pte page is pinned iff its being
1514 attached to a pinned pagetable. */
1515 static inline void xen_alloc_ptpage(struct mm_struct *mm, unsigned long pfn,
1518 bool pinned = PagePinned(virt_to_page(mm->pgd));
1520 trace_xen_mmu_alloc_ptpage(mm, pfn, level, pinned);
1523 struct page *page = pfn_to_page(pfn);
1525 SetPagePinned(page);
1527 if (!PageHighMem(page)) {
1530 __set_pfn_prot(pfn, PAGE_KERNEL_RO);
1532 if (level == PT_PTE && USE_SPLIT_PTLOCKS)
1533 __pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
1535 xen_mc_issue(PARAVIRT_LAZY_MMU);
1537 /* make sure there are no stray mappings of
1539 kmap_flush_unused();
1544 static void xen_alloc_pte(struct mm_struct *mm, unsigned long pfn)
1546 xen_alloc_ptpage(mm, pfn, PT_PTE);
1549 static void xen_alloc_pmd(struct mm_struct *mm, unsigned long pfn)
1551 xen_alloc_ptpage(mm, pfn, PT_PMD);
1554 /* This should never happen until we're OK to use struct page */
1555 static inline void xen_release_ptpage(unsigned long pfn, unsigned level)
1557 struct page *page = pfn_to_page(pfn);
1558 bool pinned = PagePinned(page);
1560 trace_xen_mmu_release_ptpage(pfn, level, pinned);
1563 if (!PageHighMem(page)) {
1566 if (level == PT_PTE && USE_SPLIT_PTLOCKS)
1567 __pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
1569 __set_pfn_prot(pfn, PAGE_KERNEL);
1571 xen_mc_issue(PARAVIRT_LAZY_MMU);
1573 ClearPagePinned(page);
1577 static void xen_release_pte(unsigned long pfn)
1579 xen_release_ptpage(pfn, PT_PTE);
1582 static void xen_release_pmd(unsigned long pfn)
1584 xen_release_ptpage(pfn, PT_PMD);
1587 #if PAGETABLE_LEVELS == 4
1588 static void xen_alloc_pud(struct mm_struct *mm, unsigned long pfn)
1590 xen_alloc_ptpage(mm, pfn, PT_PUD);
1593 static void xen_release_pud(unsigned long pfn)
1595 xen_release_ptpage(pfn, PT_PUD);
1599 void __init xen_reserve_top(void)
1601 #ifdef CONFIG_X86_32
1602 unsigned long top = HYPERVISOR_VIRT_START;
1603 struct xen_platform_parameters pp;
1605 if (HYPERVISOR_xen_version(XENVER_platform_parameters, &pp) == 0)
1606 top = pp.virt_start;
1608 reserve_top_address(-top);
1609 #endif /* CONFIG_X86_32 */
1613 * Like __va(), but returns address in the kernel mapping (which is
1614 * all we have until the physical memory mapping has been set up.
1616 static void *__ka(phys_addr_t paddr)
1618 #ifdef CONFIG_X86_64
1619 return (void *)(paddr + __START_KERNEL_map);
1625 /* Convert a machine address to physical address */
1626 static unsigned long m2p(phys_addr_t maddr)
1630 maddr &= PTE_PFN_MASK;
1631 paddr = mfn_to_pfn(maddr >> PAGE_SHIFT) << PAGE_SHIFT;
1636 /* Convert a machine address to kernel virtual */
1637 static void *m2v(phys_addr_t maddr)
1639 return __ka(m2p(maddr));
1642 /* Set the page permissions on an identity-mapped pages */
1643 static void set_page_prot(void *addr, pgprot_t prot)
1645 unsigned long pfn = __pa(addr) >> PAGE_SHIFT;
1646 pte_t pte = pfn_pte(pfn, prot);
1648 if (HYPERVISOR_update_va_mapping((unsigned long)addr, pte, 0))
1652 static void __init xen_map_identity_early(pmd_t *pmd, unsigned long max_pfn)
1654 unsigned pmdidx, pteidx;
1658 level1_ident_pgt = extend_brk(sizeof(pte_t) * LEVEL1_IDENT_ENTRIES,
1663 for (pmdidx = 0; pmdidx < PTRS_PER_PMD && pfn < max_pfn; pmdidx++) {
1666 /* Reuse or allocate a page of ptes */
1667 if (pmd_present(pmd[pmdidx]))
1668 pte_page = m2v(pmd[pmdidx].pmd);
1670 /* Check for free pte pages */
1671 if (ident_pte == LEVEL1_IDENT_ENTRIES)
1674 pte_page = &level1_ident_pgt[ident_pte];
1675 ident_pte += PTRS_PER_PTE;
1677 pmd[pmdidx] = __pmd(__pa(pte_page) | _PAGE_TABLE);
1680 /* Install mappings */
1681 for (pteidx = 0; pteidx < PTRS_PER_PTE; pteidx++, pfn++) {
1684 #ifdef CONFIG_X86_32
1685 if (pfn > max_pfn_mapped)
1686 max_pfn_mapped = pfn;
1689 if (!pte_none(pte_page[pteidx]))
1692 pte = pfn_pte(pfn, PAGE_KERNEL_EXEC);
1693 pte_page[pteidx] = pte;
1697 for (pteidx = 0; pteidx < ident_pte; pteidx += PTRS_PER_PTE)
1698 set_page_prot(&level1_ident_pgt[pteidx], PAGE_KERNEL_RO);
1700 set_page_prot(pmd, PAGE_KERNEL_RO);
1703 void __init xen_setup_machphys_mapping(void)
1705 struct xen_machphys_mapping mapping;
1707 if (HYPERVISOR_memory_op(XENMEM_machphys_mapping, &mapping) == 0) {
1708 machine_to_phys_mapping = (unsigned long *)mapping.v_start;
1709 machine_to_phys_nr = mapping.max_mfn + 1;
1711 machine_to_phys_nr = MACH2PHYS_NR_ENTRIES;
1713 #ifdef CONFIG_X86_32
1714 WARN_ON((machine_to_phys_mapping + (machine_to_phys_nr - 1))
1715 < machine_to_phys_mapping);
1719 #ifdef CONFIG_X86_64
1720 static void convert_pfn_mfn(void *v)
1725 /* All levels are converted the same way, so just treat them
1727 for (i = 0; i < PTRS_PER_PTE; i++)
1728 pte[i] = xen_make_pte(pte[i].pte);
1732 * Set up the initial kernel pagetable.
1734 * We can construct this by grafting the Xen provided pagetable into
1735 * head_64.S's preconstructed pagetables. We copy the Xen L2's into
1736 * level2_ident_pgt, level2_kernel_pgt and level2_fixmap_pgt. This
1737 * means that only the kernel has a physical mapping to start with -
1738 * but that's enough to get __va working. We need to fill in the rest
1739 * of the physical mapping once some sort of allocator has been set
1742 pgd_t * __init xen_setup_kernel_pagetable(pgd_t *pgd,
1743 unsigned long max_pfn)
1748 /* max_pfn_mapped is the last pfn mapped in the initial memory
1749 * mappings. Considering that on Xen after the kernel mappings we
1750 * have the mappings of some pages that don't exist in pfn space, we
1751 * set max_pfn_mapped to the last real pfn mapped. */
1752 max_pfn_mapped = PFN_DOWN(__pa(xen_start_info->mfn_list));
1754 /* Zap identity mapping */
1755 init_level4_pgt[0] = __pgd(0);
1757 /* Pre-constructed entries are in pfn, so convert to mfn */
1758 convert_pfn_mfn(init_level4_pgt);
1759 convert_pfn_mfn(level3_ident_pgt);
1760 convert_pfn_mfn(level3_kernel_pgt);
1762 l3 = m2v(pgd[pgd_index(__START_KERNEL_map)].pgd);
1763 l2 = m2v(l3[pud_index(__START_KERNEL_map)].pud);
1765 memcpy(level2_ident_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1766 memcpy(level2_kernel_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1768 l3 = m2v(pgd[pgd_index(__START_KERNEL_map + PMD_SIZE)].pgd);
1769 l2 = m2v(l3[pud_index(__START_KERNEL_map + PMD_SIZE)].pud);
1770 memcpy(level2_fixmap_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1772 /* Set up identity map */
1773 xen_map_identity_early(level2_ident_pgt, max_pfn);
1775 /* Make pagetable pieces RO */
1776 set_page_prot(init_level4_pgt, PAGE_KERNEL_RO);
1777 set_page_prot(level3_ident_pgt, PAGE_KERNEL_RO);
1778 set_page_prot(level3_kernel_pgt, PAGE_KERNEL_RO);
1779 set_page_prot(level3_user_vsyscall, PAGE_KERNEL_RO);
1780 set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1781 set_page_prot(level2_fixmap_pgt, PAGE_KERNEL_RO);
1783 /* Pin down new L4 */
1784 pin_pagetable_pfn(MMUEXT_PIN_L4_TABLE,
1785 PFN_DOWN(__pa_symbol(init_level4_pgt)));
1787 /* Unpin Xen-provided one */
1788 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1791 pgd = init_level4_pgt;
1794 * At this stage there can be no user pgd, and no page
1795 * structure to attach it to, so make sure we just set kernel
1799 __xen_write_cr3(true, __pa(pgd));
1800 xen_mc_issue(PARAVIRT_LAZY_CPU);
1802 memblock_reserve(__pa(xen_start_info->pt_base),
1803 xen_start_info->nr_pt_frames * PAGE_SIZE);
1807 #else /* !CONFIG_X86_64 */
1808 static RESERVE_BRK_ARRAY(pmd_t, initial_kernel_pmd, PTRS_PER_PMD);
1809 static RESERVE_BRK_ARRAY(pmd_t, swapper_kernel_pmd, PTRS_PER_PMD);
1811 static void __init xen_write_cr3_init(unsigned long cr3)
1813 unsigned long pfn = PFN_DOWN(__pa(swapper_pg_dir));
1815 BUG_ON(read_cr3() != __pa(initial_page_table));
1816 BUG_ON(cr3 != __pa(swapper_pg_dir));
1819 * We are switching to swapper_pg_dir for the first time (from
1820 * initial_page_table) and therefore need to mark that page
1821 * read-only and then pin it.
1823 * Xen disallows sharing of kernel PMDs for PAE
1824 * guests. Therefore we must copy the kernel PMD from
1825 * initial_page_table into a new kernel PMD to be used in
1828 swapper_kernel_pmd =
1829 extend_brk(sizeof(pmd_t) * PTRS_PER_PMD, PAGE_SIZE);
1830 memcpy(swapper_kernel_pmd, initial_kernel_pmd,
1831 sizeof(pmd_t) * PTRS_PER_PMD);
1832 swapper_pg_dir[KERNEL_PGD_BOUNDARY] =
1833 __pgd(__pa(swapper_kernel_pmd) | _PAGE_PRESENT);
1834 set_page_prot(swapper_kernel_pmd, PAGE_KERNEL_RO);
1836 set_page_prot(swapper_pg_dir, PAGE_KERNEL_RO);
1838 pin_pagetable_pfn(MMUEXT_PIN_L3_TABLE, pfn);
1840 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE,
1841 PFN_DOWN(__pa(initial_page_table)));
1842 set_page_prot(initial_page_table, PAGE_KERNEL);
1843 set_page_prot(initial_kernel_pmd, PAGE_KERNEL);
1845 pv_mmu_ops.write_cr3 = &xen_write_cr3;
1848 pgd_t * __init xen_setup_kernel_pagetable(pgd_t *pgd,
1849 unsigned long max_pfn)
1853 initial_kernel_pmd =
1854 extend_brk(sizeof(pmd_t) * PTRS_PER_PMD, PAGE_SIZE);
1856 max_pfn_mapped = PFN_DOWN(__pa(xen_start_info->pt_base) +
1857 xen_start_info->nr_pt_frames * PAGE_SIZE +
1860 kernel_pmd = m2v(pgd[KERNEL_PGD_BOUNDARY].pgd);
1861 memcpy(initial_kernel_pmd, kernel_pmd, sizeof(pmd_t) * PTRS_PER_PMD);
1863 xen_map_identity_early(initial_kernel_pmd, max_pfn);
1865 memcpy(initial_page_table, pgd, sizeof(pgd_t) * PTRS_PER_PGD);
1866 initial_page_table[KERNEL_PGD_BOUNDARY] =
1867 __pgd(__pa(initial_kernel_pmd) | _PAGE_PRESENT);
1869 set_page_prot(initial_kernel_pmd, PAGE_KERNEL_RO);
1870 set_page_prot(initial_page_table, PAGE_KERNEL_RO);
1871 set_page_prot(empty_zero_page, PAGE_KERNEL_RO);
1873 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1875 pin_pagetable_pfn(MMUEXT_PIN_L3_TABLE,
1876 PFN_DOWN(__pa(initial_page_table)));
1877 xen_write_cr3(__pa(initial_page_table));
1879 memblock_reserve(__pa(xen_start_info->pt_base),
1880 xen_start_info->nr_pt_frames * PAGE_SIZE);
1882 return initial_page_table;
1884 #endif /* CONFIG_X86_64 */
1886 static unsigned char dummy_mapping[PAGE_SIZE] __page_aligned_bss;
1888 static void xen_set_fixmap(unsigned idx, phys_addr_t phys, pgprot_t prot)
1892 phys >>= PAGE_SHIFT;
1895 case FIX_BTMAP_END ... FIX_BTMAP_BEGIN:
1896 #ifdef CONFIG_X86_F00F_BUG
1899 #ifdef CONFIG_X86_32
1902 # ifdef CONFIG_HIGHMEM
1903 case FIX_KMAP_BEGIN ... FIX_KMAP_END:
1906 case VSYSCALL_LAST_PAGE ... VSYSCALL_FIRST_PAGE:
1909 case FIX_TEXT_POKE0:
1910 case FIX_TEXT_POKE1:
1911 /* All local page mappings */
1912 pte = pfn_pte(phys, prot);
1915 #ifdef CONFIG_X86_LOCAL_APIC
1916 case FIX_APIC_BASE: /* maps dummy local APIC */
1917 pte = pfn_pte(PFN_DOWN(__pa(dummy_mapping)), PAGE_KERNEL);
1921 #ifdef CONFIG_X86_IO_APIC
1922 case FIX_IO_APIC_BASE_0 ... FIX_IO_APIC_BASE_END:
1924 * We just don't map the IO APIC - all access is via
1925 * hypercalls. Keep the address in the pte for reference.
1927 pte = pfn_pte(PFN_DOWN(__pa(dummy_mapping)), PAGE_KERNEL);
1931 case FIX_PARAVIRT_BOOTMAP:
1932 /* This is an MFN, but it isn't an IO mapping from the
1934 pte = mfn_pte(phys, prot);
1938 /* By default, set_fixmap is used for hardware mappings */
1939 pte = mfn_pte(phys, __pgprot(pgprot_val(prot) | _PAGE_IOMAP));
1943 __native_set_fixmap(idx, pte);
1945 #ifdef CONFIG_X86_64
1946 /* Replicate changes to map the vsyscall page into the user
1947 pagetable vsyscall mapping. */
1948 if ((idx >= VSYSCALL_LAST_PAGE && idx <= VSYSCALL_FIRST_PAGE) ||
1950 unsigned long vaddr = __fix_to_virt(idx);
1951 set_pte_vaddr_pud(level3_user_vsyscall, vaddr, pte);
1956 static void __init xen_post_allocator_init(void)
1958 pv_mmu_ops.set_pte = xen_set_pte;
1959 pv_mmu_ops.set_pmd = xen_set_pmd;
1960 pv_mmu_ops.set_pud = xen_set_pud;
1961 #if PAGETABLE_LEVELS == 4
1962 pv_mmu_ops.set_pgd = xen_set_pgd;
1965 /* This will work as long as patching hasn't happened yet
1966 (which it hasn't) */
1967 pv_mmu_ops.alloc_pte = xen_alloc_pte;
1968 pv_mmu_ops.alloc_pmd = xen_alloc_pmd;
1969 pv_mmu_ops.release_pte = xen_release_pte;
1970 pv_mmu_ops.release_pmd = xen_release_pmd;
1971 #if PAGETABLE_LEVELS == 4
1972 pv_mmu_ops.alloc_pud = xen_alloc_pud;
1973 pv_mmu_ops.release_pud = xen_release_pud;
1976 #ifdef CONFIG_X86_64
1977 SetPagePinned(virt_to_page(level3_user_vsyscall));
1979 xen_mark_init_mm_pinned();
1982 static void xen_leave_lazy_mmu(void)
1986 paravirt_leave_lazy_mmu();
1990 static const struct pv_mmu_ops xen_mmu_ops __initconst = {
1991 .read_cr2 = xen_read_cr2,
1992 .write_cr2 = xen_write_cr2,
1994 .read_cr3 = xen_read_cr3,
1995 #ifdef CONFIG_X86_32
1996 .write_cr3 = xen_write_cr3_init,
1998 .write_cr3 = xen_write_cr3,
2001 .flush_tlb_user = xen_flush_tlb,
2002 .flush_tlb_kernel = xen_flush_tlb,
2003 .flush_tlb_single = xen_flush_tlb_single,
2004 .flush_tlb_others = xen_flush_tlb_others,
2006 .pte_update = paravirt_nop,
2007 .pte_update_defer = paravirt_nop,
2009 .pgd_alloc = xen_pgd_alloc,
2010 .pgd_free = xen_pgd_free,
2012 .alloc_pte = xen_alloc_pte_init,
2013 .release_pte = xen_release_pte_init,
2014 .alloc_pmd = xen_alloc_pmd_init,
2015 .release_pmd = xen_release_pmd_init,
2017 .set_pte = xen_set_pte_init,
2018 .set_pte_at = xen_set_pte_at,
2019 .set_pmd = xen_set_pmd_hyper,
2021 .ptep_modify_prot_start = __ptep_modify_prot_start,
2022 .ptep_modify_prot_commit = __ptep_modify_prot_commit,
2024 .pte_val = PV_CALLEE_SAVE(xen_pte_val),
2025 .pgd_val = PV_CALLEE_SAVE(xen_pgd_val),
2027 .make_pte = PV_CALLEE_SAVE(xen_make_pte),
2028 .make_pgd = PV_CALLEE_SAVE(xen_make_pgd),
2030 #ifdef CONFIG_X86_PAE
2031 .set_pte_atomic = xen_set_pte_atomic,
2032 .pte_clear = xen_pte_clear,
2033 .pmd_clear = xen_pmd_clear,
2034 #endif /* CONFIG_X86_PAE */
2035 .set_pud = xen_set_pud_hyper,
2037 .make_pmd = PV_CALLEE_SAVE(xen_make_pmd),
2038 .pmd_val = PV_CALLEE_SAVE(xen_pmd_val),
2040 #if PAGETABLE_LEVELS == 4
2041 .pud_val = PV_CALLEE_SAVE(xen_pud_val),
2042 .make_pud = PV_CALLEE_SAVE(xen_make_pud),
2043 .set_pgd = xen_set_pgd_hyper,
2045 .alloc_pud = xen_alloc_pmd_init,
2046 .release_pud = xen_release_pmd_init,
2047 #endif /* PAGETABLE_LEVELS == 4 */
2049 .activate_mm = xen_activate_mm,
2050 .dup_mmap = xen_dup_mmap,
2051 .exit_mmap = xen_exit_mmap,
2054 .enter = paravirt_enter_lazy_mmu,
2055 .leave = xen_leave_lazy_mmu,
2058 .set_fixmap = xen_set_fixmap,
2061 void __init xen_init_mmu_ops(void)
2063 x86_init.mapping.pagetable_reserve = xen_mapping_pagetable_reserve;
2064 x86_init.paging.pagetable_setup_start = xen_pagetable_setup_start;
2065 x86_init.paging.pagetable_setup_done = xen_pagetable_setup_done;
2066 pv_mmu_ops = xen_mmu_ops;
2068 memset(dummy_mapping, 0xff, PAGE_SIZE);
2071 /* Protected by xen_reservation_lock. */
2072 #define MAX_CONTIG_ORDER 9 /* 2MB */
2073 static unsigned long discontig_frames[1<<MAX_CONTIG_ORDER];
2075 #define VOID_PTE (mfn_pte(0, __pgprot(0)))
2076 static void xen_zap_pfn_range(unsigned long vaddr, unsigned int order,
2077 unsigned long *in_frames,
2078 unsigned long *out_frames)
2081 struct multicall_space mcs;
2084 for (i = 0; i < (1UL<<order); i++, vaddr += PAGE_SIZE) {
2085 mcs = __xen_mc_entry(0);
2088 in_frames[i] = virt_to_mfn(vaddr);
2090 MULTI_update_va_mapping(mcs.mc, vaddr, VOID_PTE, 0);
2091 __set_phys_to_machine(virt_to_pfn(vaddr), INVALID_P2M_ENTRY);
2094 out_frames[i] = virt_to_pfn(vaddr);
2100 * Update the pfn-to-mfn mappings for a virtual address range, either to
2101 * point to an array of mfns, or contiguously from a single starting
2104 static void xen_remap_exchanged_ptes(unsigned long vaddr, int order,
2105 unsigned long *mfns,
2106 unsigned long first_mfn)
2113 limit = 1u << order;
2114 for (i = 0; i < limit; i++, vaddr += PAGE_SIZE) {
2115 struct multicall_space mcs;
2118 mcs = __xen_mc_entry(0);
2122 mfn = first_mfn + i;
2124 if (i < (limit - 1))
2128 flags = UVMF_INVLPG | UVMF_ALL;
2130 flags = UVMF_TLB_FLUSH | UVMF_ALL;
2133 MULTI_update_va_mapping(mcs.mc, vaddr,
2134 mfn_pte(mfn, PAGE_KERNEL), flags);
2136 set_phys_to_machine(virt_to_pfn(vaddr), mfn);
2143 * Perform the hypercall to exchange a region of our pfns to point to
2144 * memory with the required contiguous alignment. Takes the pfns as
2145 * input, and populates mfns as output.
2147 * Returns a success code indicating whether the hypervisor was able to
2148 * satisfy the request or not.
2150 static int xen_exchange_memory(unsigned long extents_in, unsigned int order_in,
2151 unsigned long *pfns_in,
2152 unsigned long extents_out,
2153 unsigned int order_out,
2154 unsigned long *mfns_out,
2155 unsigned int address_bits)
2160 struct xen_memory_exchange exchange = {
2162 .nr_extents = extents_in,
2163 .extent_order = order_in,
2164 .extent_start = pfns_in,
2168 .nr_extents = extents_out,
2169 .extent_order = order_out,
2170 .extent_start = mfns_out,
2171 .address_bits = address_bits,
2176 BUG_ON(extents_in << order_in != extents_out << order_out);
2178 rc = HYPERVISOR_memory_op(XENMEM_exchange, &exchange);
2179 success = (exchange.nr_exchanged == extents_in);
2181 BUG_ON(!success && ((exchange.nr_exchanged != 0) || (rc == 0)));
2182 BUG_ON(success && (rc != 0));
2187 int xen_create_contiguous_region(unsigned long vstart, unsigned int order,
2188 unsigned int address_bits)
2190 unsigned long *in_frames = discontig_frames, out_frame;
2191 unsigned long flags;
2195 * Currently an auto-translated guest will not perform I/O, nor will
2196 * it require PAE page directories below 4GB. Therefore any calls to
2197 * this function are redundant and can be ignored.
2200 if (xen_feature(XENFEAT_auto_translated_physmap))
2203 if (unlikely(order > MAX_CONTIG_ORDER))
2206 memset((void *) vstart, 0, PAGE_SIZE << order);
2208 spin_lock_irqsave(&xen_reservation_lock, flags);
2210 /* 1. Zap current PTEs, remembering MFNs. */
2211 xen_zap_pfn_range(vstart, order, in_frames, NULL);
2213 /* 2. Get a new contiguous memory extent. */
2214 out_frame = virt_to_pfn(vstart);
2215 success = xen_exchange_memory(1UL << order, 0, in_frames,
2216 1, order, &out_frame,
2219 /* 3. Map the new extent in place of old pages. */
2221 xen_remap_exchanged_ptes(vstart, order, NULL, out_frame);
2223 xen_remap_exchanged_ptes(vstart, order, in_frames, 0);
2225 spin_unlock_irqrestore(&xen_reservation_lock, flags);
2227 return success ? 0 : -ENOMEM;
2229 EXPORT_SYMBOL_GPL(xen_create_contiguous_region);
2231 void xen_destroy_contiguous_region(unsigned long vstart, unsigned int order)
2233 unsigned long *out_frames = discontig_frames, in_frame;
2234 unsigned long flags;
2237 if (xen_feature(XENFEAT_auto_translated_physmap))
2240 if (unlikely(order > MAX_CONTIG_ORDER))
2243 memset((void *) vstart, 0, PAGE_SIZE << order);
2245 spin_lock_irqsave(&xen_reservation_lock, flags);
2247 /* 1. Find start MFN of contiguous extent. */
2248 in_frame = virt_to_mfn(vstart);
2250 /* 2. Zap current PTEs. */
2251 xen_zap_pfn_range(vstart, order, NULL, out_frames);
2253 /* 3. Do the exchange for non-contiguous MFNs. */
2254 success = xen_exchange_memory(1, order, &in_frame, 1UL << order,
2257 /* 4. Map new pages in place of old pages. */
2259 xen_remap_exchanged_ptes(vstart, order, out_frames, 0);
2261 xen_remap_exchanged_ptes(vstart, order, NULL, in_frame);
2263 spin_unlock_irqrestore(&xen_reservation_lock, flags);
2265 EXPORT_SYMBOL_GPL(xen_destroy_contiguous_region);
2267 #ifdef CONFIG_XEN_PVHVM
2268 static void xen_hvm_exit_mmap(struct mm_struct *mm)
2270 struct xen_hvm_pagetable_dying a;
2273 a.domid = DOMID_SELF;
2274 a.gpa = __pa(mm->pgd);
2275 rc = HYPERVISOR_hvm_op(HVMOP_pagetable_dying, &a);
2276 WARN_ON_ONCE(rc < 0);
2279 static int is_pagetable_dying_supported(void)
2281 struct xen_hvm_pagetable_dying a;
2284 a.domid = DOMID_SELF;
2286 rc = HYPERVISOR_hvm_op(HVMOP_pagetable_dying, &a);
2288 printk(KERN_DEBUG "HVMOP_pagetable_dying not supported\n");
2294 void __init xen_hvm_init_mmu_ops(void)
2296 if (is_pagetable_dying_supported())
2297 pv_mmu_ops.exit_mmap = xen_hvm_exit_mmap;
2301 #define REMAP_BATCH_SIZE 16
2306 struct mmu_update *mmu_update;
2309 static int remap_area_mfn_pte_fn(pte_t *ptep, pgtable_t token,
2310 unsigned long addr, void *data)
2312 struct remap_data *rmd = data;
2313 pte_t pte = pte_mkspecial(pfn_pte(rmd->mfn++, rmd->prot));
2315 rmd->mmu_update->ptr = virt_to_machine(ptep).maddr;
2316 rmd->mmu_update->val = pte_val_ma(pte);
2322 int xen_remap_domain_mfn_range(struct vm_area_struct *vma,
2324 unsigned long mfn, int nr,
2325 pgprot_t prot, unsigned domid)
2327 struct remap_data rmd;
2328 struct mmu_update mmu_update[REMAP_BATCH_SIZE];
2330 unsigned long range;
2333 prot = __pgprot(pgprot_val(prot) | _PAGE_IOMAP);
2335 BUG_ON(!((vma->vm_flags & (VM_PFNMAP | VM_RESERVED | VM_IO)) ==
2336 (VM_PFNMAP | VM_RESERVED | VM_IO)));
2342 batch = min(REMAP_BATCH_SIZE, nr);
2343 range = (unsigned long)batch << PAGE_SHIFT;
2345 rmd.mmu_update = mmu_update;
2346 err = apply_to_page_range(vma->vm_mm, addr, range,
2347 remap_area_mfn_pte_fn, &rmd);
2352 if (HYPERVISOR_mmu_update(mmu_update, batch, NULL, domid) < 0)
2366 EXPORT_SYMBOL_GPL(xen_remap_domain_mfn_range);